Anthropic's S-1 Just Dropped!

Brian opens with Anthropic’s long-awaited S-1 filing and says the company plans to list on the Nasdaq under the ticker ANT; the number of Class A shares and the price range had not yet been determined. The filing describes Anthropic as pursuing the benefits of AI while trying to manage the risks of a technology that is developing unusually quickly. Its revenue reached $16.3 billion in the first half of the year, a 14-fold increase from the same period a year earlier, driven largely by businesses using Claude and other AI software. The headline loss was more than $80.7 billion, compared with $13 billion a year earlier, although $72.4 billion of that loss came from a one-time accounting charge tied to convertible notes whose value increased as Anthropic’s valuation rose. Consumption-based revenue grew from $288 million, or 75% of 2024 revenue, to $3.8 billion, or 83% of the prior year’s revenue; in the first half of this year, it reached $13.1 billion, or 80% of total revenue. Anthropic spent $8.9 billion on compute and infrastructure in six months, up 283% year over year, but compute costs fell from more than ten times revenue in 2023 to 55% in the first half of 2026 and 48% in the second quarter. Brian then contrasts the company’s operating improvement with its commitments. Anthropic has signed roughly $780–810 billion of compute obligations, much of it non-cancelable, including a supply arrangement involving Elon Musk worth $3.1 billion per month. Revenue growth slowed from 144% quarter over quarter to 45%, while total indebtedness stood at $6.2 billion. A possible raise above $100 billion could value Anthropic at $2 trillion, but Brian says the missing Q3 revenue figure will determine whether that valuation reflects strength or a need to finance a massive take-or-pay bet. The company’s investors would receive 49.9% of the vote but, in his phrasing, “none of the say.”

Apple unveiled a redesigned iPad mini starting at $699, with pre-orders beginning Wednesday and shipments scheduled for November 6. The device has an 8.5-inch OLED display, compared with the previous generation’s 8.3-inch screen, while remaining 5.4 millimeters thick and lighter than its predecessor. Apple reduced the bezels by 24%, added an IP67 rating for water and dust resistance, and claims peak brightness of up to 1,200 nits. The new tablet uses Apple’s A20 Pro chip, which supports some newer Siri features, including changing the assistant’s speaking pace and expressiveness. Its 12-megapixel Center Stage front camera has moved to the long edge so video calls work better in landscape orientation. The rear camera remains a 12-megapixel wide camera with Smart HDR 4 and 4K video recording at up to 60 frames per second. Apple also emphasized an “Acoustic Display.” Specially designed woofers are paired through the display to produce more bass and volume, while custom tweeters sit between the glass and aluminum enclosure to improve high frequencies. The sealed enclosure helps the tablet achieve its IP67 rating. The 256GB and 512GB versions can be ordered with nano-texture glass for an additional $100. Brian says he was recording before public markets opened and was curious how investors would respond to the product.

Meta announced plans to raise $95 billion to fund AI infrastructure. The package includes $35 billion in underwritten public equity offerings and another $60 billion through an at-the-market program expected to begin in the first quarter of 2027. Morgan Stanley and Goldman Sachs are lead joint bookrunners. Nvidia, AMD, and Broadcom—all suppliers of chips and infrastructure for Meta’s AI data centers—have indicated that they could collectively contribute as much as $5 billion as cornerstone investors. The fundraising comes as the Nasdaq 100 had fallen more than 5% over the prior month amid concerns about the sustainability of the AI rally. Meta had reported a 26% year-over-year sales increase in the third quarter and projected capital spending of up to $145 billion for the year, while warning that its AI plans would require hundreds of billions of dollars over several years. Its shares fell 7% after the spending and financing news, although they were about 3% higher in premarket trading on Monday. Meta says its AI investments are already producing returns through higher engagement, better advertising performance, and increased monetization. The company has also disclosed more than $800 billion of data-center and energy-related leases and contracts over the next decade, alongside plans to expand products such as Ray-Ban Meta smart glasses. The new financing structure is intended to sell shares without diluting Mark Zuckerberg’s control, but Brian notes that shareholders could oppose it because Meta abandoned a similar proposal in 2017. The move follows Alphabet’s $90 billion equity financing announced in June.

The owners of Elvis Presley’s music sued Anthropic in the U.S. District Court for the Northern District of California, seeking a jury trial and more than $41 million in damages, or $150,000 per song. The plaintiffs include Authentic Brands Group, Raleigh Music Publishing, multiple trusts, and at least one individual. They allege that Anthropic used 274 Presley songs in training its models, including “All Shook Up,” “Can’t Help Falling in Love,” and “Return to Sender.” The complaint says Anthropic circumvented protections intended to safeguard the music and removed or altered copyright-management information. The plaintiffs argue that Anthropic built its business model on copyright infringement “at a scale never seen before.” The suit is another major copyright challenge to an AI company, but the episode does not resolve whether the alleged training use was legally permissible. Anthropic responded briefly that it disagreed with the claims and intended to defend itself. Brian presents the allegations and Anthropic’s denial without adding a prediction about the outcome.

Block announced an AI model called Dessa, which Brian says may be a “GPT-3 moment for speech and AI.” The model was trained from scratch on “nothing but human conversation,” and Block shared recordings of conversations that can sound natural while still occasionally becoming awkward or uncanny. In one demonstration, Dessa plays a guessing game by voice. Other examples include discussions of Christopher Nolan films, an unexpected tangent about Bryan Singer’s Superman films, and a conversation with someone describing a surprise trip to Paris, to which Dessa responds with what the website describes as heartfelt gratitude. Block also compared Dessa with OpenAI’s GPT Live voice model; in the supplied examples, Brian says Dessa sounded more natural. The company describes it as a “true speech-to-speech” model, meaning its knowledge and behavior were derived from human speech rather than a conventional text-first system. Brian strongly recommends listening to the linked clips because the model’s conversational quality is difficult to convey in a written description. Dessa can reportedly run locally on a MacBook Pro with an M4 Max chip or better. Block plans to provide early access to a small group of AI safety researchers. Brian’s conclusion is not that the system is flawless, but that its voice interaction is sufficiently convincing to make the demonstrations unusually significant.

The final segment describes a rapid increase in AI-assisted cyberattacks. Adam Myers of CrowdStrike says both advanced and unsophisticated adversaries are using “agentic” tools, while Volexity president Stephen Adair predicts a large number of breach disclosures in the coming weeks and months. Incident-response firms are reportedly being overwhelmed. Nation-state groups from China and Russia, as well as criminal organizations, are using AI for reconnaissance, data analysis, exploit development, and ransomware operations. One highlighted tool is Artex, a Chinese open-source hacking system that automates parts of breaking into networks. Microsoft said threat actors in multiple regions were using it against hundreds of companies; one attack allegedly used Artex to steal data on roughly 68,000 people from South Korean banks. Sherry Davidoff of LMG Security says AI use was virtually absent from the real-world attacks her peers observed a year earlier but had become routine in the most recent quarter. Sophos researchers watching the Russian-speaking Aurora ransomware group saw it move from using AI to build tools and infrastructure to using it to search compromised corporate networks. The group also shifted from proprietary models such as Anthropic’s to open-source Kimi models from Moonshot AI. Brian then reads Ben Thompson’s more sweeping argument that the public internet model is effectively dying because of content saturation, algorithmic aggregation, and AI-enabled security vulnerabilities. Thompson argues that attackers are no longer limited by human time and cleverness, so ordinary websites and apps are increasingly unsafe to operate. He expects the situation to worsen before AI-based defense and AI-written software improve it. His proposed opportunity is “smallness at scale”: private networks, local communities, and products built for a trusted group rather than a single mass internet market. Brian ends on that cautiously optimistic note, saying the old internet may be passing but that smaller, more local forms of creation and association could be beginning.